Multi-objective Optimization in Autonomous DrivingBasedon Reinforcement Learning

With autonomous driving technology playing an increasingly important role in intelligent transportation systems, how to improve ride comfort while ensuring safety has become an urgent challenge. This paper proposes a multi-objective optimization approach for autonomous driving based on reinforcement learning. By designing a multi-objective reward function that integrates rewards based on position, speed, direction, and acceleration, the method aims to balance driving efficiency and ride comfort. The Proximal Policy Optimization (PPO) algorithm is employed for training on the high-fidelity simulation platform MetaDrive, and experiments in multiple scenarios verify the effectiveness of the proposed approach. The experimental results show that as the comfort penalty coefficient in the reward function changes, the success rates for left turns, straight driving, and right turns exhibit a non-linear trend of first increasing then decreasing, with the best performance achieved when the parameter value is 0.001. This fully demonstrates the critical impact of parameter selection on the performance of autonomous driving strategies. It provides an optimization solution for reinforcement learning-based autonomous driving decision-making that balances safety and ride comfort, and offers a reference for subsequent related research.